ultra

Domain toolkit and rigor for single-cell / spatial-omics graph analysis — community detection on cell or gene graphs, spatial neighbor graphs, ligand-receptor inference, and graph neural networks. Use when a task involves single-cell or spatial transcriptomics, clustering cells/genes, detecting modules or communities in a biological graph, testing whether that structure is real, or ligand-receptor / cell-cell communication. Read it before writing code that would otherwise reach for a hand-rolled surrogate (e.g. plain Louvain, a synthetic ligand-receptor score, a home-made SBM null).

Domain toolkit and rigor for geospatial and ecological analysis — vector/raster IO, CRS handling, spatial statistics, point patterns, landscape metrics, terrain/hydrology, remote-sensing indices, movement tracks, habitat connectivity, and species distribution modelling. Use for shapefiles, GeoPackage, GeoTIFF, NetCDF/COG rasters, reprojection, clustered-vs-random spatial tests, patch metrics, NDVI/EVI, zonal statistics, GPS telemetry, least-cost corridors, or occurrence prediction. Read before hand-rolling spatial weights, Moran's I, landscape metrics, NDVI loops, or resistance-surface nulls.

How to render a chart, plot, graph, figure, heatmap, or diagram in the code-execution sandbox so it looks calm and native to Ultra. Covers chart-type choice (form follows data), the shipped Ultra matplotlib style + brand color palette, whisper-quiet chrome, direct labeling over legends, colorblind-safe color used only to encode meaning, light-on-white 300-DPI export to /outputs, and the ONE offline-safe way to make diagrams (render to an image — mermaid/graphviz do not work in the sandbox). Use when generating any data visualization or diagram as an image via code execution. For where files go, captioning, and report structure, use scientific-reporting.

Render a small INTERACTIVE chart inline in the chat conversation (hover tooltips, legend, light/dark aware) by emitting a fenced ```chart block containing a JSON spec — no code execution, no image file. Use for a quick line/bar/area/scatter/pie of small tabular data (a few series, up to ~1000 rows) that you already have in the conversation and want the reader to hover and read. For publication-quality, complex, or scientific figures, or anything needing computation, use plots-and-diagrams (matplotlib, static image via code execution) instead.

Turn a 3D/4D medical imaging volume (CT, MRI, fMRI, DICOM series, or tractography) into clean, VLM-readable 2D anatomical slices, then have the vision-reasoner inspect them. Use when the user uploads or asks about a NIfTI (.nii/.nii.gz), DICOM, or tractography (.trk/.tck) volume — head CT, brain MRI, fMRI, diffusion/tractography, or similar — and wants it described, assessed, or visually inspected.

Run RareSpot prairie-dog and burrow detection on field images and present per-detection reliability for an ecologist — tiled YOLOv5 inference plus perturbation-consensus stability scoring, an energy-adaptive spectral mask, and an EXIF survey map. Use when the user asks to detect, count, or quantify prairie dogs or burrows in one or more images, run RareSpot, or survey colony activity — i.e. running ecological image detection, not writing it up.

Rigor protocol for the analysis itself — uncertainty quantification, convergence checks, classification thresholds, literature grounding, and reproducibility records. Use when designing or running a simulation, numerical experiment, or parameter study, when computing estimates (exponents, rates, fits, spectra, regime classifications), or before making any quantitative claim — i.e. while producing the numbers, not while writing them up.

Report contract and artifact hygiene for the write-up — required sections (methods, uncertainties, verification, limitations, references), figure standards, subagent attribution, and durable-output discipline. Use when writing a scientific report, a results summary for a researcher, or producing durable deliverables (figures, CSVs, code, markdown) — i.e. while communicating results, not while computing them.